ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/libecb/ecb.h
(Generate patch)

Comparing libecb/ecb.h (file contents):
Revision 1.164 by root, Mon Apr 20 20:06:30 2015 UTC vs.
Revision 1.202 by root, Wed Mar 23 09:58:06 2022 UTC

1/* 1/*
2 * libecb - http://software.schmorp.de/pkg/libecb 2 * libecb - http://software.schmorp.de/pkg/libecb
3 * 3 *
4 * Copyright (©) 2009-2015 Marc Alexander Lehmann <libecb@schmorp.de> 4 * Copyright (©) 2009-2015,2018-2021 Marc Alexander Lehmann <libecb@schmorp.de>
5 * Copyright (©) 2011 Emanuele Giaquinta 5 * Copyright (©) 2011 Emanuele Giaquinta
6 * All rights reserved. 6 * All rights reserved.
7 * 7 *
8 * Redistribution and use in source and binary forms, with or without modifica- 8 * Redistribution and use in source and binary forms, with or without modifica-
9 * tion, are permitted provided that the following conditions are met: 9 * tion, are permitted provided that the following conditions are met:
… …
40 40
41#ifndef ECB_H 41#ifndef ECB_H
42#define ECB_H 42#define ECB_H
43 43
44/* 16 bits major, 16 bits minor */ 44/* 16 bits major, 16 bits minor */
45#define ECB_VERSION 0x00010004 45#define ECB_VERSION 0x0001000b
46 46
47#ifdef _WIN32 47#include <string.h> /* for memcpy */
48
49#if defined (_WIN32) && !defined (__MINGW32__)
48 typedef signed char int8_t; 50 typedef signed char int8_t;
49 typedef unsigned char uint8_t; 51 typedef unsigned char uint8_t;
52 typedef signed char int_fast8_t;
53 typedef unsigned char uint_fast8_t;
50 typedef signed short int16_t; 54 typedef signed short int16_t;
51 typedef unsigned short uint16_t; 55 typedef unsigned short uint16_t;
56 typedef signed int int_fast16_t;
57 typedef unsigned int uint_fast16_t;
52 typedef signed int int32_t; 58 typedef signed int int32_t;
53 typedef unsigned int uint32_t; 59 typedef unsigned int uint32_t;
60 typedef signed int int_fast32_t;
61 typedef unsigned int uint_fast32_t;
54 #if __GNUC__ 62 #if __GNUC__
55 typedef signed long long int64_t; 63 typedef signed long long int64_t;
56 typedef unsigned long long uint64_t; 64 typedef unsigned long long uint64_t;
57 #else /* _MSC_VER || __BORLANDC__ */ 65 #else /* _MSC_VER || __BORLANDC__ */
58 typedef signed __int64 int64_t; 66 typedef signed __int64 int64_t;
59 typedef unsigned __int64 uint64_t; 67 typedef unsigned __int64 uint64_t;
60 #endif 68 #endif
69 typedef int64_t int_fast64_t;
70 typedef uint64_t uint_fast64_t;
61 #ifdef _WIN64 71 #ifdef _WIN64
62 #define ECB_PTRSIZE 8 72 #define ECB_PTRSIZE 8
63 typedef uint64_t uintptr_t; 73 typedef uint64_t uintptr_t;
64 typedef int64_t intptr_t; 74 typedef int64_t intptr_t;
65 #else 75 #else
… …
67 typedef uint32_t uintptr_t; 77 typedef uint32_t uintptr_t;
68 typedef int32_t intptr_t; 78 typedef int32_t intptr_t;
69 #endif 79 #endif
70#else 80#else
71 #include <inttypes.h> 81 #include <inttypes.h>
72 #if UINTMAX_MAX > 0xffffffffU 82 #if (defined INTPTR_MAX ? INTPTR_MAX : ULONG_MAX) > 0xffffffffU
73 #define ECB_PTRSIZE 8 83 #define ECB_PTRSIZE 8
74 #else 84 #else
75 #define ECB_PTRSIZE 4 85 #define ECB_PTRSIZE 4
76 #endif 86 #endif
77#endif 87#endif
78 88
79#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__) 89#define ECB_GCC_AMD64 (__amd64 || __amd64__ || __x86_64 || __x86_64__)
80#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64) 90#define ECB_MSVC_AMD64 (_M_AMD64 || _M_X64)
91
92#ifndef ECB_OPTIMIZE_SIZE
93 #if __OPTIMIZE_SIZE__
94 #define ECB_OPTIMIZE_SIZE 1
95 #else
96 #define ECB_OPTIMIZE_SIZE 0
97 #endif
98#endif
81 99
82/* work around x32 idiocy by defining proper macros */ 100/* work around x32 idiocy by defining proper macros */
83#if ECB_GCC_AMD64 || ECB_MSVC_AMD64 101#if ECB_GCC_AMD64 || ECB_MSVC_AMD64
84 #if _ILP32 102 #if _ILP32
85 #define ECB_AMD64_X32 1 103 #define ECB_AMD64_X32 1
86 #else 104 #else
87 #define ECB_AMD64 1 105 #define ECB_AMD64 1
88 #endif 106 #endif
107#endif
108
109#if ECB_PTRSIZE >= 8 || ECB_AMD64_X32
110 #define ECB_64BIT_NATIVE 1
111#else
112 #define ECB_64BIT_NATIVE 0
89#endif 113#endif
90 114
91/* many compilers define _GNUC_ to some versions but then only implement 115/* many compilers define _GNUC_ to some versions but then only implement
92 * what their idiot authors think are the "more important" extensions, 116 * what their idiot authors think are the "more important" extensions,
93 * causing enormous grief in return for some better fake benchmark numbers. 117 * causing enormous grief in return for some better fake benchmark numbers.
… …
115 #define ECB_CLANG_EXTENSION(x) 0 139 #define ECB_CLANG_EXTENSION(x) 0
116#endif 140#endif
117 141
118#define ECB_CPP (__cplusplus+0) 142#define ECB_CPP (__cplusplus+0)
119#define ECB_CPP11 (__cplusplus >= 201103L) 143#define ECB_CPP11 (__cplusplus >= 201103L)
144#define ECB_CPP14 (__cplusplus >= 201402L)
145#define ECB_CPP17 (__cplusplus >= 201703L)
120 146
121#if ECB_CPP 147#if ECB_CPP
122 #define ECB_C 0 148 #define ECB_C 0
123 #define ECB_STDC_VERSION 0 149 #define ECB_STDC_VERSION 0
124#else 150#else
… …
126 #define ECB_STDC_VERSION __STDC_VERSION__ 152 #define ECB_STDC_VERSION __STDC_VERSION__
127#endif 153#endif
128 154
129#define ECB_C99 (ECB_STDC_VERSION >= 199901L) 155#define ECB_C99 (ECB_STDC_VERSION >= 199901L)
130#define ECB_C11 (ECB_STDC_VERSION >= 201112L) 156#define ECB_C11 (ECB_STDC_VERSION >= 201112L)
157#define ECB_C17 (ECB_STDC_VERSION >= 201710L)
131 158
132#if ECB_CPP 159#if ECB_CPP
133 #define ECB_EXTERN_C extern "C" 160 #define ECB_EXTERN_C extern "C"
134 #define ECB_EXTERN_C_BEG ECB_EXTERN_C { 161 #define ECB_EXTERN_C_BEG ECB_EXTERN_C {
135 #define ECB_EXTERN_C_END } 162 #define ECB_EXTERN_C_END }
… …
150 177
151#if ECB_NO_SMP 178#if ECB_NO_SMP
152 #define ECB_MEMORY_FENCE do { } while (0) 179 #define ECB_MEMORY_FENCE do { } while (0)
153#endif 180#endif
154 181
182/* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/compiler_ref/compiler_builtins.html */
183#if __xlC__ && ECB_CPP
184 #include <builtins.h>
185#endif
186
187#if 1400 <= _MSC_VER
188 #include <intrin.h> /* fence functions _ReadBarrier, also bit search functions _BitScanReverse */
189#endif
190
155#ifndef ECB_MEMORY_FENCE 191#ifndef ECB_MEMORY_FENCE
156 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 192 #if ECB_GCC_VERSION(2,5) || defined __INTEL_COMPILER || (__llvm__ && __GNUC__) || __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
193 #define ECB_MEMORY_FENCE_RELAXED __asm__ __volatile__ ("" : : : "memory")
157 #if __i386 || __i386__ 194 #if __i386 || __i386__
158 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory") 195 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("lock; orb $0, -1(%%esp)" : : : "memory")
159 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 196 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
160 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 197 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
161 #elif ECB_GCC_AMD64 198 #elif ECB_GCC_AMD64
162 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory") 199 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mfence" : : : "memory")
163 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory") 200 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("" : : : "memory")
164 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("") 201 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("" : : : "memory")
165 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__ 202 #elif __powerpc__ || __ppc__ || __powerpc64__ || __ppc64__
166 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory") 203 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("sync" : : : "memory")
204 #elif defined __ARM_ARCH_2__ \
205 || defined __ARM_ARCH_3__ || defined __ARM_ARCH_3M__ \
206 || defined __ARM_ARCH_4__ || defined __ARM_ARCH_4T__ \
207 || defined __ARM_ARCH_5__ || defined __ARM_ARCH_5E__ \
208 || defined __ARM_ARCH_5T__ || defined __ARM_ARCH_5TE__ \
209 || defined __ARM_ARCH_5TEJ__
210 /* should not need any, unless running old code on newer cpu - arm doesn't support that */
167 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \ 211 #elif defined __ARM_ARCH_6__ || defined __ARM_ARCH_6J__ \
168 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ 212 || defined __ARM_ARCH_6K__ || defined __ARM_ARCH_6ZK__ \
213 || defined __ARM_ARCH_6T2__
169 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory") 214 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("mcr p15,0,%0,c7,c10,5" : : "r" (0) : "memory")
170 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \ 215 #elif defined __ARM_ARCH_7__ || defined __ARM_ARCH_7A__ \
171 || defined __ARM_ARCH_7M__ || defined __ARM_ARCH_7R__ 216 || defined __ARM_ARCH_7R__ || defined __ARM_ARCH_7M__
172 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory") 217 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb" : : : "memory")
173 #elif __aarch64__ 218 #elif __aarch64__
174 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory") 219 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("dmb ish" : : : "memory")
175 #elif (__sparc || __sparc__) && !__sparcv8 220 #elif (__sparc || __sparc__) && !(__sparc_v8__ || defined __sparcv8)
176 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory") 221 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad | #StoreStore | #StoreLoad" : : : "memory")
177 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory") 222 #define ECB_MEMORY_FENCE_ACQUIRE __asm__ __volatile__ ("membar #LoadStore | #LoadLoad" : : : "memory")
178 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore") 223 #define ECB_MEMORY_FENCE_RELEASE __asm__ __volatile__ ("membar #LoadStore | #StoreStore")
179 #elif defined __s390__ || defined __s390x__ 224 #elif defined __s390__ || defined __s390x__
180 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory") 225 #define ECB_MEMORY_FENCE __asm__ __volatile__ ("bcr 15,0" : : : "memory")
… …
203 #if ECB_GCC_VERSION(4,7) 248 #if ECB_GCC_VERSION(4,7)
204 /* see comment below (stdatomic.h) about the C11 memory model. */ 249 /* see comment below (stdatomic.h) about the C11 memory model. */
205 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST) 250 #define ECB_MEMORY_FENCE __atomic_thread_fence (__ATOMIC_SEQ_CST)
206 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE) 251 #define ECB_MEMORY_FENCE_ACQUIRE __atomic_thread_fence (__ATOMIC_ACQUIRE)
207 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE) 252 #define ECB_MEMORY_FENCE_RELEASE __atomic_thread_fence (__ATOMIC_RELEASE)
253 #undef ECB_MEMORY_FENCE_RELAXED
254 #define ECB_MEMORY_FENCE_RELAXED __atomic_thread_fence (__ATOMIC_RELAXED)
208 255
209 #elif ECB_CLANG_EXTENSION(c_atomic) 256 #elif ECB_CLANG_EXTENSION(c_atomic)
210 /* see comment below (stdatomic.h) about the C11 memory model. */ 257 /* see comment below (stdatomic.h) about the C11 memory model. */
211 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST) 258 #define ECB_MEMORY_FENCE __c11_atomic_thread_fence (__ATOMIC_SEQ_CST)
212 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE) 259 #define ECB_MEMORY_FENCE_ACQUIRE __c11_atomic_thread_fence (__ATOMIC_ACQUIRE)
213 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE) 260 #define ECB_MEMORY_FENCE_RELEASE __c11_atomic_thread_fence (__ATOMIC_RELEASE)
261 #undef ECB_MEMORY_FENCE_RELAXED
262 #define ECB_MEMORY_FENCE_RELAXED __c11_atomic_thread_fence (__ATOMIC_RELAXED)
214 263
215 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__ 264 #elif ECB_GCC_VERSION(4,4) || defined __INTEL_COMPILER || defined __clang__
216 #define ECB_MEMORY_FENCE __sync_synchronize () 265 #define ECB_MEMORY_FENCE __sync_synchronize ()
217 #elif _MSC_VER >= 1500 /* VC++ 2008 */ 266 #elif _MSC_VER >= 1500 /* VC++ 2008 */
218 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */ 267 /* apparently, microsoft broke all the memory barrier stuff in Visual Studio 2008... */
… …
228 #elif defined _WIN32 277 #elif defined _WIN32
229 #include <WinNT.h> 278 #include <WinNT.h>
230 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */ 279 #define ECB_MEMORY_FENCE MemoryBarrier () /* actually just xchg on x86... scary */
231 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110 280 #elif __SUNPRO_C >= 0x5110 || __SUNPRO_CC >= 0x5110
232 #include <mbarrier.h> 281 #include <mbarrier.h>
233 #define ECB_MEMORY_FENCE __machine_rw_barrier () 282 #define ECB_MEMORY_FENCE __machine_rw_barrier ()
234 #define ECB_MEMORY_FENCE_ACQUIRE __machine_r_barrier () 283 #define ECB_MEMORY_FENCE_ACQUIRE __machine_acq_barrier ()
235 #define ECB_MEMORY_FENCE_RELEASE __machine_w_barrier () 284 #define ECB_MEMORY_FENCE_RELEASE __machine_rel_barrier ()
285 #define ECB_MEMORY_FENCE_RELAXED __compiler_barrier ()
236 #elif __xlC__ 286 #elif __xlC__
237 #define ECB_MEMORY_FENCE __sync () 287 #define ECB_MEMORY_FENCE __sync ()
238 #endif 288 #endif
239#endif 289#endif
240 290
241#ifndef ECB_MEMORY_FENCE 291#ifndef ECB_MEMORY_FENCE
242 #if ECB_C11 && !defined __STDC_NO_ATOMICS__ 292 #if ECB_C11 && !defined __STDC_NO_ATOMICS__
243 /* we assume that these memory fences work on all variables/all memory accesses, */ 293 /* we assume that these memory fences work on all variables/all memory accesses, */
244 /* not just C11 atomics and atomic accesses */ 294 /* not just C11 atomics and atomic accesses */
245 #include <stdatomic.h> 295 #include <stdatomic.h>
246 /* Unfortunately, neither gcc 4.7 nor clang 3.1 generate any instructions for */
247 /* any fence other than seq_cst, which isn't very efficient for us. */
248 /* Why that is, we don't know - either the C11 memory model is quite useless */
249 /* for most usages, or gcc and clang have a bug */
250 /* I *currently* lean towards the latter, and inefficiently implement */
251 /* all three of ecb's fences as a seq_cst fence */
252 /* Update, gcc-4.8 generates mfence for all c++ fences, but nothing */
253 /* for all __atomic_thread_fence's except seq_cst */
254 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst) 296 #define ECB_MEMORY_FENCE atomic_thread_fence (memory_order_seq_cst)
297 #define ECB_MEMORY_FENCE_ACQUIRE atomic_thread_fence (memory_order_acquire)
298 #define ECB_MEMORY_FENCE_RELEASE atomic_thread_fence (memory_order_release)
255 #endif 299 #endif
256#endif 300#endif
257 301
258#ifndef ECB_MEMORY_FENCE 302#ifndef ECB_MEMORY_FENCE
259 #if !ECB_AVOID_PTHREADS 303 #if !ECB_AVOID_PTHREADS
… …
279 323
280#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE 324#if !defined ECB_MEMORY_FENCE_RELEASE && defined ECB_MEMORY_FENCE
281 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE 325 #define ECB_MEMORY_FENCE_RELEASE ECB_MEMORY_FENCE
282#endif 326#endif
283 327
328#if !defined ECB_MEMORY_FENCE_RELAXED && defined ECB_MEMORY_FENCE
329 #define ECB_MEMORY_FENCE_RELAXED ECB_MEMORY_FENCE /* very heavy-handed */
330#endif
331
284/*****************************************************************************/ 332/*****************************************************************************/
285 333
286#if ECB_CPP 334#if ECB_CPP
287 #define ecb_inline static inline 335 #define ecb_inline static inline
288#elif ECB_GCC_VERSION(2,5) 336#elif ECB_GCC_VERSION(2,5)
… …
371#define ecb_unused ecb_attribute ((__unused__)) 419#define ecb_unused ecb_attribute ((__unused__))
372#define ecb_const ecb_attribute ((__const__)) 420#define ecb_const ecb_attribute ((__const__))
373#define ecb_pure ecb_attribute ((__pure__)) 421#define ecb_pure ecb_attribute ((__pure__))
374 422
375#if ECB_C11 || __IBMC_NORETURN 423#if ECB_C11 || __IBMC_NORETURN
376 /* http://pic.dhe.ibm.com/infocenter/compbg/v121v141/topic/com.ibm.xlcpp121.bg.doc/language_ref/noreturn.html */ 424 /* http://www-01.ibm.com/support/knowledgecenter/SSGH3R_13.1.0/com.ibm.xlcpp131.aix.doc/language_ref/noreturn.html */
377 #define ecb_noreturn _Noreturn 425 #define ecb_noreturn _Noreturn
378#elif ECB_CPP11 426#elif ECB_CPP11
379 #define ecb_noreturn [[noreturn]] 427 #define ecb_noreturn [[noreturn]]
380#elif _MSC_VER >= 1200 428#elif _MSC_VER >= 1200
381 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */ 429 /* http://msdn.microsoft.com/en-us/library/k6ktzx3s.aspx */
… …
418#else 466#else
419 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x); 467 ecb_function_ ecb_const int ecb_ctz32 (uint32_t x);
420 ecb_function_ ecb_const int 468 ecb_function_ ecb_const int
421 ecb_ctz32 (uint32_t x) 469 ecb_ctz32 (uint32_t x)
422 { 470 {
471#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
472 unsigned long r;
473 _BitScanForward (&r, x);
474 return (int)r;
475#else
423 int r = 0; 476 int r = 0;
424 477
425 x &= ~x + 1; /* this isolates the lowest bit */ 478 x &= ~x + 1; /* this isolates the lowest bit */
426 479
427#if ECB_branchless_on_i386 480#if ECB_branchless_on_i386
… …
437 if (x & 0xff00ff00) r += 8; 490 if (x & 0xff00ff00) r += 8;
438 if (x & 0xffff0000) r += 16; 491 if (x & 0xffff0000) r += 16;
439#endif 492#endif
440 493
441 return r; 494 return r;
495#endif
442 } 496 }
443 497
444 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x); 498 ecb_function_ ecb_const int ecb_ctz64 (uint64_t x);
445 ecb_function_ ecb_const int 499 ecb_function_ ecb_const int
446 ecb_ctz64 (uint64_t x) 500 ecb_ctz64 (uint64_t x)
447 { 501 {
502#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
503 unsigned long r;
504 _BitScanForward64 (&r, x);
505 return (int)r;
506#else
448 int shift = x & 0xffffffffU ? 0 : 32; 507 int shift = x & 0xffffffff ? 0 : 32;
449 return ecb_ctz32 (x >> shift) + shift; 508 return ecb_ctz32 (x >> shift) + shift;
509#endif
450 } 510 }
451 511
452 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x); 512 ecb_function_ ecb_const int ecb_popcount32 (uint32_t x);
453 ecb_function_ ecb_const int 513 ecb_function_ ecb_const int
454 ecb_popcount32 (uint32_t x) 514 ecb_popcount32 (uint32_t x)
… …
462 } 522 }
463 523
464 ecb_function_ ecb_const int ecb_ld32 (uint32_t x); 524 ecb_function_ ecb_const int ecb_ld32 (uint32_t x);
465 ecb_function_ ecb_const int ecb_ld32 (uint32_t x) 525 ecb_function_ ecb_const int ecb_ld32 (uint32_t x)
466 { 526 {
527#if 1400 <= _MSC_VER && (_M_IX86 || _M_X64 || _M_IA64 || _M_ARM)
528 unsigned long r;
529 _BitScanReverse (&r, x);
530 return (int)r;
531#else
467 int r = 0; 532 int r = 0;
468 533
469 if (x >> 16) { x >>= 16; r += 16; } 534 if (x >> 16) { x >>= 16; r += 16; }
470 if (x >> 8) { x >>= 8; r += 8; } 535 if (x >> 8) { x >>= 8; r += 8; }
471 if (x >> 4) { x >>= 4; r += 4; } 536 if (x >> 4) { x >>= 4; r += 4; }
472 if (x >> 2) { x >>= 2; r += 2; } 537 if (x >> 2) { x >>= 2; r += 2; }
473 if (x >> 1) { r += 1; } 538 if (x >> 1) { r += 1; }
474 539
475 return r; 540 return r;
541#endif
476 } 542 }
477 543
478 ecb_function_ ecb_const int ecb_ld64 (uint64_t x); 544 ecb_function_ ecb_const int ecb_ld64 (uint64_t x);
479 ecb_function_ ecb_const int ecb_ld64 (uint64_t x) 545 ecb_function_ ecb_const int ecb_ld64 (uint64_t x)
480 { 546 {
547#if 1400 <= _MSC_VER && (_M_X64 || _M_IA64 || _M_ARM)
548 unsigned long r;
549 _BitScanReverse64 (&r, x);
550 return (int)r;
551#else
481 int r = 0; 552 int r = 0;
482 553
483 if (x >> 32) { x >>= 32; r += 32; } 554 if (x >> 32) { x >>= 32; r += 32; }
484 555
485 return r + ecb_ld32 (x); 556 return r + ecb_ld32 (x);
557#endif
486 } 558 }
487#endif 559#endif
488 560
489ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x); 561ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x);
490ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); } 562ecb_function_ ecb_const ecb_bool ecb_is_pot32 (uint32_t x) { return !(x & (x - 1)); }
… …
537ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count); 609ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count);
538ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count); 610ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count);
539ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count); 611ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count);
540ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count); 612ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count);
541 613
542ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> ( 8 - count)) | (x << count); } 614ecb_inline ecb_const uint8_t ecb_rotl8 (uint8_t x, unsigned int count) { return (x >> (-count & 7)) | (x << (count & 7)); }
543ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << ( 8 - count)) | (x >> count); } 615ecb_inline ecb_const uint8_t ecb_rotr8 (uint8_t x, unsigned int count) { return (x << (-count & 7)) | (x >> (count & 7)); }
544ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (16 - count)) | (x << count); } 616ecb_inline ecb_const uint16_t ecb_rotl16 (uint16_t x, unsigned int count) { return (x >> (-count & 15)) | (x << (count & 15)); }
545ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (16 - count)) | (x >> count); } 617ecb_inline ecb_const uint16_t ecb_rotr16 (uint16_t x, unsigned int count) { return (x << (-count & 15)) | (x >> (count & 15)); }
546ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (32 - count)) | (x << count); } 618ecb_inline ecb_const uint32_t ecb_rotl32 (uint32_t x, unsigned int count) { return (x >> (-count & 31)) | (x << (count & 31)); }
547ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (32 - count)) | (x >> count); } 619ecb_inline ecb_const uint32_t ecb_rotr32 (uint32_t x, unsigned int count) { return (x << (-count & 31)) | (x >> (count & 31)); }
548ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (64 - count)) | (x << count); } 620ecb_inline ecb_const uint64_t ecb_rotl64 (uint64_t x, unsigned int count) { return (x >> (-count & 63)) | (x << (count & 63)); }
549ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (64 - count)) | (x >> count); } 621ecb_inline ecb_const uint64_t ecb_rotr64 (uint64_t x, unsigned int count) { return (x << (-count & 63)) | (x >> (count & 63)); }
622
623#if ECB_CPP
624
625inline uint8_t ecb_ctz (uint8_t v) { return ecb_ctz32 (v); }
626inline uint16_t ecb_ctz (uint16_t v) { return ecb_ctz32 (v); }
627inline uint32_t ecb_ctz (uint32_t v) { return ecb_ctz32 (v); }
628inline uint64_t ecb_ctz (uint64_t v) { return ecb_ctz64 (v); }
629
630inline bool ecb_is_pot (uint8_t v) { return ecb_is_pot32 (v); }
631inline bool ecb_is_pot (uint16_t v) { return ecb_is_pot32 (v); }
632inline bool ecb_is_pot (uint32_t v) { return ecb_is_pot32 (v); }
633inline bool ecb_is_pot (uint64_t v) { return ecb_is_pot64 (v); }
634
635inline int ecb_ld (uint8_t v) { return ecb_ld32 (v); }
636inline int ecb_ld (uint16_t v) { return ecb_ld32 (v); }
637inline int ecb_ld (uint32_t v) { return ecb_ld32 (v); }
638inline int ecb_ld (uint64_t v) { return ecb_ld64 (v); }
639
640inline int ecb_popcount (uint8_t v) { return ecb_popcount32 (v); }
641inline int ecb_popcount (uint16_t v) { return ecb_popcount32 (v); }
642inline int ecb_popcount (uint32_t v) { return ecb_popcount32 (v); }
643inline int ecb_popcount (uint64_t v) { return ecb_popcount64 (v); }
644
645inline uint8_t ecb_bitrev (uint8_t v) { return ecb_bitrev8 (v); }
646inline uint16_t ecb_bitrev (uint16_t v) { return ecb_bitrev16 (v); }
647inline uint32_t ecb_bitrev (uint32_t v) { return ecb_bitrev32 (v); }
648
649inline uint8_t ecb_rotl (uint8_t v, unsigned int count) { return ecb_rotl8 (v, count); }
650inline uint16_t ecb_rotl (uint16_t v, unsigned int count) { return ecb_rotl16 (v, count); }
651inline uint32_t ecb_rotl (uint32_t v, unsigned int count) { return ecb_rotl32 (v, count); }
652inline uint64_t ecb_rotl (uint64_t v, unsigned int count) { return ecb_rotl64 (v, count); }
653
654inline uint8_t ecb_rotr (uint8_t v, unsigned int count) { return ecb_rotr8 (v, count); }
655inline uint16_t ecb_rotr (uint16_t v, unsigned int count) { return ecb_rotr16 (v, count); }
656inline uint32_t ecb_rotr (uint32_t v, unsigned int count) { return ecb_rotr32 (v, count); }
657inline uint64_t ecb_rotr (uint64_t v, unsigned int count) { return ecb_rotr64 (v, count); }
658
659#endif
550 660
551#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64)) 661#if ECB_GCC_VERSION(4,3) || (ECB_CLANG_BUILTIN(__builtin_bswap32) && ECB_CLANG_BUILTIN(__builtin_bswap64))
552 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16) 662 #if ECB_GCC_VERSION(4,8) || ECB_CLANG_BUILTIN(__builtin_bswap16)
553 #define ecb_bswap16(x) __builtin_bswap16 (x) 663 #define ecb_bswap16(x) __builtin_bswap16 (x)
554 #else 664 #else
… …
593#endif 703#endif
594 704
595/* try to tell the compiler that some condition is definitely true */ 705/* try to tell the compiler that some condition is definitely true */
596#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0 706#define ecb_assume(cond) if (!(cond)) ecb_unreachable (); else 0
597 707
598ecb_inline ecb_const unsigned char ecb_byteorder_helper (void); 708ecb_inline ecb_const uint32_t ecb_byteorder_helper (void);
599ecb_inline ecb_const unsigned char 709ecb_inline ecb_const uint32_t
600ecb_byteorder_helper (void) 710ecb_byteorder_helper (void)
601{ 711{
602 /* the union code still generates code under pressure in gcc, */ 712 /* the union code still generates code under pressure in gcc, */
603 /* but less than using pointers, and always seems to */ 713 /* but less than using pointers, and always seems to */
604 /* successfully return a constant. */ 714 /* successfully return a constant. */
605 /* the reason why we have this horrible preprocessor mess */ 715 /* the reason why we have this horrible preprocessor mess */
606 /* is to avoid it in all cases, at least on common architectures */ 716 /* is to avoid it in all cases, at least on common architectures */
607 /* or when using a recent enough gcc version (>= 4.6) */ 717 /* or when using a recent enough gcc version (>= 4.6) */
608#if ((__i386 || __i386__) && !__VOS__) || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64
609 return 0x44;
610#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ 718#if (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__) \
719 || ((__i386 || __i386__ || _M_IX86 || ECB_GCC_AMD64 || ECB_MSVC_AMD64) && !__VOS__)
720 #define ECB_LITTLE_ENDIAN 1
611 return 0x44; 721 return 0x44332211;
612#elif __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__ 722#elif (defined __BYTE_ORDER__ && __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) \
723 || ((__AARCH64EB__ || __MIPSEB__ || __ARMEB__) && !__VOS__)
724 #define ECB_BIG_ENDIAN 1
613 return 0x11; 725 return 0x11223344;
614#else 726#else
615 union 727 union
616 { 728 {
729 uint8_t c[4];
617 uint32_t i; 730 uint32_t u;
618 uint8_t c;
619 } u = { 0x11223344 }; 731 } u = { 0x11, 0x22, 0x33, 0x44 };
620 return u.c; 732 return u.u;
621#endif 733#endif
622} 734}
623 735
624ecb_inline ecb_const ecb_bool ecb_big_endian (void); 736ecb_inline ecb_const ecb_bool ecb_big_endian (void);
625ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11; } 737ecb_inline ecb_const ecb_bool ecb_big_endian (void) { return ecb_byteorder_helper () == 0x11223344; }
626ecb_inline ecb_const ecb_bool ecb_little_endian (void); 738ecb_inline ecb_const ecb_bool ecb_little_endian (void);
627ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44; } 739ecb_inline ecb_const ecb_bool ecb_little_endian (void) { return ecb_byteorder_helper () == 0x44332211; }
740
741/*****************************************************************************/
742/* unaligned load/store */
743
744ecb_inline uint_fast16_t ecb_be_u16_to_host (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
745ecb_inline uint_fast32_t ecb_be_u32_to_host (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
746ecb_inline uint_fast64_t ecb_be_u64_to_host (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
747
748ecb_inline uint_fast16_t ecb_le_u16_to_host (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
749ecb_inline uint_fast32_t ecb_le_u32_to_host (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
750ecb_inline uint_fast64_t ecb_le_u64_to_host (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
751
752ecb_inline uint_fast16_t ecb_peek_u16_u (const void *ptr) { uint16_t v; memcpy (&v, ptr, sizeof (v)); return v; }
753ecb_inline uint_fast32_t ecb_peek_u32_u (const void *ptr) { uint32_t v; memcpy (&v, ptr, sizeof (v)); return v; }
754ecb_inline uint_fast64_t ecb_peek_u64_u (const void *ptr) { uint64_t v; memcpy (&v, ptr, sizeof (v)); return v; }
755
756ecb_inline uint_fast16_t ecb_peek_be_u16_u (const void *ptr) { return ecb_be_u16_to_host (ecb_peek_u16_u (ptr)); }
757ecb_inline uint_fast32_t ecb_peek_be_u32_u (const void *ptr) { return ecb_be_u32_to_host (ecb_peek_u32_u (ptr)); }
758ecb_inline uint_fast64_t ecb_peek_be_u64_u (const void *ptr) { return ecb_be_u64_to_host (ecb_peek_u64_u (ptr)); }
759
760ecb_inline uint_fast16_t ecb_peek_le_u16_u (const void *ptr) { return ecb_le_u16_to_host (ecb_peek_u16_u (ptr)); }
761ecb_inline uint_fast32_t ecb_peek_le_u32_u (const void *ptr) { return ecb_le_u32_to_host (ecb_peek_u32_u (ptr)); }
762ecb_inline uint_fast64_t ecb_peek_le_u64_u (const void *ptr) { return ecb_le_u64_to_host (ecb_peek_u64_u (ptr)); }
763
764ecb_inline uint_fast16_t ecb_host_to_be_u16 (uint_fast16_t v) { return ecb_little_endian () ? ecb_bswap16 (v) : v; }
765ecb_inline uint_fast32_t ecb_host_to_be_u32 (uint_fast32_t v) { return ecb_little_endian () ? ecb_bswap32 (v) : v; }
766ecb_inline uint_fast64_t ecb_host_to_be_u64 (uint_fast64_t v) { return ecb_little_endian () ? ecb_bswap64 (v) : v; }
767
768ecb_inline uint_fast16_t ecb_host_to_le_u16 (uint_fast16_t v) { return ecb_big_endian () ? ecb_bswap16 (v) : v; }
769ecb_inline uint_fast32_t ecb_host_to_le_u32 (uint_fast32_t v) { return ecb_big_endian () ? ecb_bswap32 (v) : v; }
770ecb_inline uint_fast64_t ecb_host_to_le_u64 (uint_fast64_t v) { return ecb_big_endian () ? ecb_bswap64 (v) : v; }
771
772ecb_inline void ecb_poke_u16_u (void *ptr, uint16_t v) { memcpy (ptr, &v, sizeof (v)); }
773ecb_inline void ecb_poke_u32_u (void *ptr, uint32_t v) { memcpy (ptr, &v, sizeof (v)); }
774ecb_inline void ecb_poke_u64_u (void *ptr, uint64_t v) { memcpy (ptr, &v, sizeof (v)); }
775
776ecb_inline void ecb_poke_be_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_be_u16 (v)); }
777ecb_inline void ecb_poke_be_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_be_u32 (v)); }
778ecb_inline void ecb_poke_be_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_be_u64 (v)); }
779
780ecb_inline void ecb_poke_le_u16_u (void *ptr, uint_fast16_t v) { ecb_poke_u16_u (ptr, ecb_host_to_le_u16 (v)); }
781ecb_inline void ecb_poke_le_u32_u (void *ptr, uint_fast32_t v) { ecb_poke_u32_u (ptr, ecb_host_to_le_u32 (v)); }
782ecb_inline void ecb_poke_le_u64_u (void *ptr, uint_fast64_t v) { ecb_poke_u64_u (ptr, ecb_host_to_le_u64 (v)); }
783
784#if ECB_CPP
785
786inline uint8_t ecb_bswap (uint8_t v) { return v; }
787inline uint16_t ecb_bswap (uint16_t v) { return ecb_bswap16 (v); }
788inline uint32_t ecb_bswap (uint32_t v) { return ecb_bswap32 (v); }
789inline uint64_t ecb_bswap (uint64_t v) { return ecb_bswap64 (v); }
790
791template<typename T> inline T ecb_be_to_host (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
792template<typename T> inline T ecb_le_to_host (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
793template<typename T> inline T ecb_peek (const void *ptr) { return *(const T *)ptr; }
794template<typename T> inline T ecb_peek_be (const void *ptr) { return ecb_be_to_host (ecb_peek <T> (ptr)); }
795template<typename T> inline T ecb_peek_le (const void *ptr) { return ecb_le_to_host (ecb_peek <T> (ptr)); }
796template<typename T> inline T ecb_peek_u (const void *ptr) { T v; memcpy (&v, ptr, sizeof (v)); return v; }
797template<typename T> inline T ecb_peek_be_u (const void *ptr) { return ecb_be_to_host (ecb_peek_u<T> (ptr)); }
798template<typename T> inline T ecb_peek_le_u (const void *ptr) { return ecb_le_to_host (ecb_peek_u<T> (ptr)); }
799
800template<typename T> inline T ecb_host_to_be (T v) { return ecb_little_endian () ? ecb_bswap (v) : v; }
801template<typename T> inline T ecb_host_to_le (T v) { return ecb_big_endian () ? ecb_bswap (v) : v; }
802template<typename T> inline void ecb_poke (void *ptr, T v) { *(T *)ptr = v; }
803template<typename T> inline void ecb_poke_be (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_be (v)); }
804template<typename T> inline void ecb_poke_le (void *ptr, T v) { return ecb_poke <T> (ptr, ecb_host_to_le (v)); }
805template<typename T> inline void ecb_poke_u (void *ptr, T v) { memcpy (ptr, &v, sizeof (v)); }
806template<typename T> inline void ecb_poke_be_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_be (v)); }
807template<typename T> inline void ecb_poke_le_u (void *ptr, T v) { return ecb_poke_u<T> (ptr, ecb_host_to_le (v)); }
808
809#endif
810
811/*****************************************************************************/
812/* pointer/integer hashing */
813
814/* based on hash by Chris Wellons, https://nullprogram.com/blog/2018/07/31/ */
815ecb_function_ uint32_t ecb_mix32 (uint32_t v);
816ecb_function_ uint32_t ecb_mix32 (uint32_t v)
817{
818 v ^= v >> 16; v *= 0x7feb352dU;
819 v ^= v >> 15; v *= 0x846ca68bU;
820 v ^= v >> 16;
821 return v;
822}
823
824ecb_function_ uint32_t ecb_unmix32 (uint32_t v);
825ecb_function_ uint32_t ecb_unmix32 (uint32_t v)
826{
827 v ^= v >> 16 ; v *= 0x43021123U;
828 v ^= v >> 15 ^ v >> 30; v *= 0x1d69e2a5U;
829 v ^= v >> 16 ;
830 return v;
831}
832
833/* based on splitmix64, by Sebastiona Vigna, https://prng.di.unimi.it/splitmix64.c */
834ecb_function_ uint64_t ecb_mix64 (uint64_t v);
835ecb_function_ uint64_t ecb_mix64 (uint64_t v)
836{
837 v ^= v >> 30; v *= 0xbf58476d1ce4e5b9U;
838 v ^= v >> 27; v *= 0x94d049bb133111ebU;
839 v ^= v >> 31;
840 return v;
841}
842
843ecb_function_ uint64_t ecb_unmix64 (uint64_t v);
844ecb_function_ uint64_t ecb_unmix64 (uint64_t v)
845{
846 v ^= v >> 31 ^ v >> 62; v *= 0x319642b2d24d8ec3U;
847 v ^= v >> 27 ^ v >> 54; v *= 0x96de1b173f119089U;
848 v ^= v >> 30 ^ v >> 60;
849 return v;
850}
851
852ecb_function_ uintptr_t ecb_ptrmix (void *p);
853ecb_function_ uintptr_t ecb_ptrmix (void *p)
854{
855 #if ECB_PTRSIZE <= 4
856 return ecb_mix32 ((uint32_t)p);
857 #else
858 return ecb_mix64 ((uint64_t)p);
859 #endif
860}
861
862ecb_function_ void *ecb_ptrunmix (uintptr_t v);
863ecb_function_ void *ecb_ptrunmix (uintptr_t v)
864{
865 #if ECB_PTRSIZE <= 4
866 return (void *)ecb_unmix32 (v);
867 #else
868 return (void *)ecb_unmix64 (v);
869 #endif
870}
871
872#if ECB_CPP
873
874template<typename T>
875inline uintptr_t ecb_ptrmix (T *p)
876{
877 return ecb_ptrmix (static_cast<void *>(p));
878}
879
880template<typename T>
881inline T *ecb_ptrunmix (uintptr_t v)
882{
883 return static_cast<T *>(ecb_ptrunmix (v));
884}
885
886#endif
887
888/*****************************************************************************/
889/* gray code */
890
891ecb_function_ uint_fast8_t ecb_gray8_encode (uint_fast8_t b) { return b ^ (b >> 1); }
892ecb_function_ uint_fast16_t ecb_gray16_encode (uint_fast16_t b) { return b ^ (b >> 1); }
893ecb_function_ uint_fast32_t ecb_gray32_encode (uint_fast32_t b) { return b ^ (b >> 1); }
894ecb_function_ uint_fast64_t ecb_gray64_encode (uint_fast64_t b) { return b ^ (b >> 1); }
895
896ecb_function_ uint8_t ecb_gray8_decode (uint8_t g)
897{
898 g = g ^ (g >> 1);
899 g = g ^ (g >> 2);
900 g = g ^ (g >> 4);
901 return g;
902}
903
904ecb_function_ uint16_t ecb_gray16_decode (uint16_t g)
905{
906 g = g ^ (g >> 1);
907 g = g ^ (g >> 2);
908 g = g ^ (g >> 4);
909 g = g ^ (g >> 8);
910 return g;
911}
912
913ecb_function_ uint32_t ecb_gray32_decode (uint32_t g)
914{
915 g = g ^ (g >> 1);
916 g = g ^ (g >> 2);
917 g = g ^ (g >> 4);
918 g = g ^ (g >> 8);
919 g = g ^ (g >> 16);
920 return g;
921}
922
923ecb_function_ uint64_t ecb_gray64_decode (uint64_t g)
924{
925 g = g ^ (g >> 1);
926 g = g ^ (g >> 2);
927 g = g ^ (g >> 4);
928 g = g ^ (g >> 8);
929 g = g ^ (g >> 16);
930 g = g ^ (g >> 32);
931 return g;
932}
933
934#if ECB_CPP
935
936ecb_function_ uint8_t ecb_gray_encode (uint8_t b) { return ecb_gray8_encode (b); }
937ecb_function_ uint16_t ecb_gray_encode (uint16_t b) { return ecb_gray16_encode (b); }
938ecb_function_ uint32_t ecb_gray_encode (uint32_t b) { return ecb_gray32_encode (b); }
939ecb_function_ uint64_t ecb_gray_encode (uint64_t b) { return ecb_gray64_encode (b); }
940
941ecb_function_ uint8_t ecb_gray_decode (uint8_t g) { return ecb_gray8_decode (g); }
942ecb_function_ uint16_t ecb_gray_decode (uint16_t g) { return ecb_gray16_decode (g); }
943ecb_function_ uint32_t ecb_gray_decode (uint32_t g) { return ecb_gray32_decode (g); }
944ecb_function_ uint64_t ecb_gray_decode (uint64_t g) { return ecb_gray64_decode (g); }
945
946#endif
947
948/*****************************************************************************/
949/* division */
628 950
629#if ECB_GCC_VERSION(3,0) || ECB_C99 951#if ECB_GCC_VERSION(3,0) || ECB_C99
952 /* C99 tightened the definition of %, so we can use a more efficient version */
630 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0)) 953 #define ecb_mod(m,n) ((m) % (n) + ((m) % (n) < 0 ? (n) : 0))
631#else 954#else
632 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n))) 955 #define ecb_mod(m,n) ((m) < 0 ? ((n) - 1 - ((-1 - (m)) % (n))) : ((m) % (n)))
633#endif 956#endif
634 957
… …
645 } 968 }
646#else 969#else
647 #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div)) 970 #define ecb_div_rd(val,div) ((val) < 0 ? - ((-(val) + (div) - 1) / (div)) : ((val) ) / (div))
648 #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div)) 971 #define ecb_div_ru(val,div) ((val) < 0 ? - ((-(val) ) / (div)) : ((val) + (div) - 1) / (div))
649#endif 972#endif
973
974/*****************************************************************************/
975/* array length */
650 976
651#if ecb_cplusplus_does_not_suck 977#if ecb_cplusplus_does_not_suck
652 /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */ 978 /* does not work for local types (http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2008/n2657.htm) */
653 template<typename T, int N> 979 template<typename T, int N>
654 static inline int ecb_array_length (const T (&arr)[N]) 980 static inline int ecb_array_length (const T (&arr)[N])
… …
656 return N; 982 return N;
657 } 983 }
658#else 984#else
659 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0])) 985 #define ecb_array_length(name) (sizeof (name) / sizeof (name [0]))
660#endif 986#endif
987
988/*****************************************************************************/
989/* IEEE 754-2008 half float conversions */
990
991ecb_function_ ecb_const uint32_t ecb_binary16_to_binary32 (uint32_t x);
992ecb_function_ ecb_const uint32_t
993ecb_binary16_to_binary32 (uint32_t x)
994{
995 unsigned int s = (x & 0x8000) << (31 - 15);
996 int e = (x >> 10) & 0x001f;
997 unsigned int m = x & 0x03ff;
998
999 if (ecb_expect_false (e == 31))
1000 /* infinity or NaN */
1001 e = 255 - (127 - 15);
1002 else if (ecb_expect_false (!e))
1003 {
1004 if (ecb_expect_true (!m))
1005 /* zero, handled by code below by forcing e to 0 */
1006 e = 0 - (127 - 15);
1007 else
1008 {
1009 /* subnormal, renormalise */
1010 unsigned int s = 10 - ecb_ld32 (m);
1011
1012 m = (m << s) & 0x3ff; /* mask implicit bit */
1013 e -= s - 1;
1014 }
1015 }
1016
1017 /* e and m now are normalised, or zero, (or inf or nan) */
1018 e += 127 - 15;
1019
1020 return s | (e << 23) | (m << (23 - 10));
1021}
1022
1023ecb_function_ ecb_const uint16_t ecb_binary32_to_binary16 (uint32_t x);
1024ecb_function_ ecb_const uint16_t
1025ecb_binary32_to_binary16 (uint32_t x)
1026{
1027 unsigned int s = (x >> 16) & 0x00008000; /* sign bit, the easy part */
1028 int e = ((x >> 23) & 0x000000ff) - (127 - 15); /* the desired exponent */
1029 unsigned int m = x & 0x007fffff;
1030
1031 x &= 0x7fffffff;
1032
1033 /* if it's within range of binary16 normals, use fast path */
1034 if (ecb_expect_true (0x38800000 <= x && x <= 0x477fefff))
1035 {
1036 /* mantissa round-to-even */
1037 m += 0x00000fff + ((m >> (23 - 10)) & 1);
1038
1039 /* handle overflow */
1040 if (ecb_expect_false (m >= 0x00800000))
1041 {
1042 m >>= 1;
1043 e += 1;
1044 }
1045
1046 return s | (e << 10) | (m >> (23 - 10));
1047 }
1048
1049 /* handle large numbers and infinity */
1050 if (ecb_expect_true (0x477fefff < x && x <= 0x7f800000))
1051 return s | 0x7c00;
1052
1053 /* handle zero, subnormals and small numbers */
1054 if (ecb_expect_true (x < 0x38800000))
1055 {
1056 /* zero */
1057 if (ecb_expect_true (!x))
1058 return s;
1059
1060 /* handle subnormals */
1061
1062 /* too small, will be zero */
1063 if (e < (14 - 24)) /* might not be sharp, but is good enough */
1064 return s;
1065
1066 m |= 0x00800000; /* make implicit bit explicit */
1067
1068 /* very tricky - we need to round to the nearest e (+10) bit value */
1069 {
1070 unsigned int bits = 14 - e;
1071 unsigned int half = (1 << (bits - 1)) - 1;
1072 unsigned int even = (m >> bits) & 1;
1073
1074 /* if this overflows, we will end up with a normalised number */
1075 m = (m + half + even) >> bits;
1076 }
1077
1078 return s | m;
1079 }
1080
1081 /* handle NaNs, preserve leftmost nan bits, but make sure we don't turn them into infinities */
1082 m >>= 13;
1083
1084 return s | 0x7c00 | m | !m;
1085}
1086
1087/*******************************************************************************/
1088/* fast integer to ascii */
1089
1090/*
1091 * This code is pretty complicated because it is general. The idea behind it,
1092 * however, is pretty simple: first, the number is multiplied with a scaling
1093 * factor (2**bits / 10**(digits-1)) to convert the integer into a fixed-point
1094 * number with the first digit in the upper bits.
1095 * Then this digit is converted to text and masked out. The resulting number
1096 * is then multiplied by 10, by multiplying the fixed point representation
1097 * by 5 and shifting the (binary) decimal point one to the right, so a 4.28
1098 * format becomes 5.27, 6.26 and so on.
1099 * The rest involves only advancing the pointer if we already generated a
1100 * non-zero digit, so leading zeroes are overwritten.
1101 */
1102
1103/* simply return a mask with "bits" bits set */
1104#define ecb_i2a_mask(type,bits) ((((type)1) << (bits)) - 1)
1105
1106/* oputput a single digit. maskvalue is 10**digitidx */
1107#define ecb_i2a_digit(type,bits,digitmask,maskvalue,digitidx) \
1108 if (digitmask >= maskvalue) /* constant, used to decide how many digits to generate */ \
1109 { \
1110 char digit = x >> (bits - digitidx); /* calculate the topmost digit */ \
1111 *ptr = digit + '0'; /* output it */ \
1112 nz = (digitmask == maskvalue) || nz || digit; /* first term == always output last digit */ \
1113 ptr += nz; /* output digit only if non-zero digit seen */ \
1114 x = (x & ecb_i2a_mask (type, bits - digitidx)) * 5; /* *10, but shift decimal point right */ \
1115 }
1116
1117/* convert integer to fixed point format and multiply out digits, highest first */
1118/* requires magic constants: max. digits and number of bits after the decimal point */
1119#define ecb_i2a_def(suffix,ptr,v,type,bits,digitmask,lz) \
1120ecb_inline char *ecb_i2a_ ## suffix (char *ptr, uint32_t u) \
1121{ \
1122 char nz = lz; /* non-zero digit seen? */ \
1123 /* convert to x.bits fixed-point */ \
1124 type x = u * ((ecb_i2a_mask (type, bits) + digitmask) / digitmask); \
1125 /* output up to 10 digits */ \
1126 ecb_i2a_digit (type,bits,digitmask, 1, 0); \
1127 ecb_i2a_digit (type,bits,digitmask, 10, 1); \
1128 ecb_i2a_digit (type,bits,digitmask, 100, 2); \
1129 ecb_i2a_digit (type,bits,digitmask, 1000, 3); \
1130 ecb_i2a_digit (type,bits,digitmask, 10000, 4); \
1131 ecb_i2a_digit (type,bits,digitmask, 100000, 5); \
1132 ecb_i2a_digit (type,bits,digitmask, 1000000, 6); \
1133 ecb_i2a_digit (type,bits,digitmask, 10000000, 7); \
1134 ecb_i2a_digit (type,bits,digitmask, 100000000, 8); \
1135 ecb_i2a_digit (type,bits,digitmask, 1000000000, 9); \
1136 return ptr; \
1137}
1138
1139/* predefined versions of the above, for various digits */
1140/* ecb_i2a_xN = almost N digits, limit defined by macro */
1141/* ecb_i2a_N = up to N digits, leading zeroes suppressed */
1142/* ecb_i2a_0N = exactly N digits, including leading zeroes */
1143
1144/* non-leading-zero versions, limited range */
1145#define ECB_I2A_MAX_X5 59074 /* limit for ecb_i2a_x5 */
1146#define ECB_I2A_MAX_X10 2932500665 /* limit for ecb_i2a_x10 */
1147ecb_i2a_def ( x5, ptr, v, uint32_t, 26, 10000, 0)
1148ecb_i2a_def (x10, ptr, v, uint64_t, 60, 1000000000, 0)
1149
1150/* non-leading zero versions, all digits, 4 and 9 are optimal for 32/64 bit */
1151ecb_i2a_def ( 2, ptr, v, uint32_t, 10, 10, 0)
1152ecb_i2a_def ( 3, ptr, v, uint32_t, 12, 100, 0)
1153ecb_i2a_def ( 4, ptr, v, uint32_t, 26, 1000, 0)
1154ecb_i2a_def ( 5, ptr, v, uint64_t, 30, 10000, 0)
1155ecb_i2a_def ( 6, ptr, v, uint64_t, 36, 100000, 0)
1156ecb_i2a_def ( 7, ptr, v, uint64_t, 44, 1000000, 0)
1157ecb_i2a_def ( 8, ptr, v, uint64_t, 50, 10000000, 0)
1158ecb_i2a_def ( 9, ptr, v, uint64_t, 56, 100000000, 0)
1159
1160/* leading-zero versions, all digits, 04 and 09 are optimal for 32/64 bit */
1161ecb_i2a_def (02, ptr, v, uint32_t, 10, 10, 1)
1162ecb_i2a_def (03, ptr, v, uint32_t, 12, 100, 1)
1163ecb_i2a_def (04, ptr, v, uint32_t, 26, 1000, 1)
1164ecb_i2a_def (05, ptr, v, uint64_t, 30, 10000, 1)
1165ecb_i2a_def (06, ptr, v, uint64_t, 36, 100000, 1)
1166ecb_i2a_def (07, ptr, v, uint64_t, 44, 1000000, 1)
1167ecb_i2a_def (08, ptr, v, uint64_t, 50, 10000000, 1)
1168ecb_i2a_def (09, ptr, v, uint64_t, 56, 100000000, 1)
1169
1170#define ECB_I2A_I32_DIGITS 11
1171#define ECB_I2A_U32_DIGITS 10
1172#define ECB_I2A_I64_DIGITS 20
1173#define ECB_I2A_U64_DIGITS 21
1174#define ECB_I2A_MAX_DIGITS 21
1175
1176ecb_inline char *
1177ecb_i2a_u32 (char *ptr, uint32_t u)
1178{
1179 #if ECB_64BIT_NATIVE
1180 if (ecb_expect_true (u <= ECB_I2A_MAX_X10))
1181 ptr = ecb_i2a_x10 (ptr, u);
1182 else /* x10 almost, but not fully, covers 32 bit */
1183 {
1184 uint32_t u1 = u % 1000000000;
1185 uint32_t u2 = u / 1000000000;
1186
1187 *ptr++ = u2 + '0';
1188 ptr = ecb_i2a_09 (ptr, u1);
1189 }
1190 #else
1191 if (ecb_expect_true (u <= ECB_I2A_MAX_X5))
1192 ecb_i2a_x5 (ptr, u);
1193 else if (ecb_expect_true (u <= ECB_I2A_MAX_X5 * 10000))
1194 {
1195 uint32_t u1 = u % 10000;
1196 uint32_t u2 = u / 10000;
1197
1198 ptr = ecb_i2a_x5 (ptr, u2);
1199 ptr = ecb_i2a_04 (ptr, u1);
1200 }
1201 else
1202 {
1203 uint32_t u1 = u % 10000;
1204 uint32_t ua = u / 10000;
1205 uint32_t u2 = ua % 10000;
1206 uint32_t u3 = ua / 10000;
1207
1208 ptr = ecb_i2a_2 (ptr, u3);
1209 ptr = ecb_i2a_04 (ptr, u2);
1210 ptr = ecb_i2a_04 (ptr, u1);
1211 }
1212 #endif
1213
1214 return ptr;
1215}
1216
1217ecb_inline char *
1218ecb_i2a_i32 (char *ptr, int32_t v)
1219{
1220 *ptr = '-'; ptr += v < 0;
1221 uint32_t u = v < 0 ? -(uint32_t)v : v;
1222
1223 #if ECB_64BIT_NATIVE
1224 ptr = ecb_i2a_x10 (ptr, u); /* x10 fully covers 31 bit */
1225 #else
1226 ptr = ecb_i2a_u32 (ptr, u);
1227 #endif
1228
1229 return ptr;
1230}
1231
1232ecb_inline char *
1233ecb_i2a_u64 (char *ptr, uint64_t u)
1234{
1235 #if ECB_64BIT_NATIVE
1236 if (ecb_expect_true (u <= ECB_I2A_MAX_X10))
1237 ptr = ecb_i2a_x10 (ptr, u);
1238 else if (ecb_expect_false (u <= ECB_I2A_MAX_X10 * 1000000000))
1239 {
1240 uint64_t u1 = u % 1000000000;
1241 uint64_t u2 = u / 1000000000;
1242
1243 ptr = ecb_i2a_x10 (ptr, u2);
1244 ptr = ecb_i2a_09 (ptr, u1);
1245 }
1246 else
1247 {
1248 uint64_t u1 = u % 1000000000;
1249 uint64_t ua = u / 1000000000;
1250 uint64_t u2 = ua % 1000000000;
1251 uint64_t u3 = ua / 1000000000;
1252
1253 ptr = ecb_i2a_2 (ptr, u3);
1254 ptr = ecb_i2a_09 (ptr, u2);
1255 ptr = ecb_i2a_09 (ptr, u1);
1256 }
1257 #else
1258 if (ecb_expect_true (u <= ECB_I2A_MAX_X5))
1259 ptr = ecb_i2a_x5 (ptr, u);
1260 else
1261 {
1262 uint64_t u1 = u % 10000;
1263 uint64_t u2 = u / 10000;
1264
1265 ptr = ecb_i2a_u64 (ptr, u2);
1266 ptr = ecb_i2a_04 (ptr, u1);
1267 }
1268 #endif
1269
1270 return ptr;
1271}
1272
1273ecb_inline char *
1274ecb_i2a_i64 (char *ptr, int64_t v)
1275{
1276 *ptr = '-'; ptr += v < 0;
1277 uint64_t u = v < 0 ? -(uint64_t)v : v;
1278
1279 #if ECB_64BIT_NATIVE
1280 if (ecb_expect_true (u <= ECB_I2A_MAX_X10))
1281 ptr = ecb_i2a_x10 (ptr, u);
1282 else if (ecb_expect_false (u <= ECB_I2A_MAX_X10 * 1000000000))
1283 {
1284 uint64_t u1 = u % 1000000000;
1285 uint64_t u2 = u / 1000000000;
1286
1287 ptr = ecb_i2a_x10 (ptr, u2);
1288 ptr = ecb_i2a_09 (ptr, u1);
1289 }
1290 else
1291 {
1292 uint64_t u1 = u % 1000000000;
1293 uint64_t ua = u / 1000000000;
1294 uint64_t u2 = ua % 1000000000;
1295 uint64_t u3 = ua / 1000000000;
1296
1297 /* 2**31 is 19 digits, so the top is exactly one digit */
1298 *ptr++ = u3 + '0';
1299 ptr = ecb_i2a_09 (ptr, u2);
1300 ptr = ecb_i2a_09 (ptr, u1);
1301 }
1302 #else
1303 ptr = ecb_i2a_u64 (ptr, u);
1304 #endif
1305
1306 return ptr;
1307}
661 1308
662/*******************************************************************************/ 1309/*******************************************************************************/
663/* floating point stuff, can be disabled by defining ECB_NO_LIBM */ 1310/* floating point stuff, can be disabled by defining ECB_NO_LIBM */
664 1311
665/* basically, everything uses "ieee pure-endian" floating point numbers */ 1312/* basically, everything uses "ieee pure-endian" floating point numbers */
… …
678 || defined __sh__ \ 1325 || defined __sh__ \
679 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \ 1326 || defined _M_IX86 || defined ECB_MSVC_AMD64 || defined _M_IA64 \
680 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \ 1327 || (defined __arm__ && (defined __ARM_EABI__ || defined __EABI__ || defined __VFP_FP__ || defined _WIN32_WCE || defined __ANDROID__)) \
681 || defined __aarch64__ 1328 || defined __aarch64__
682 #define ECB_STDFP 1 1329 #define ECB_STDFP 1
683 #include <string.h> /* for memcpy */
684#else 1330#else
685 #define ECB_STDFP 0 1331 #define ECB_STDFP 0
686#endif 1332#endif
687 1333
688#ifndef ECB_NO_LIBM 1334#ifndef ECB_NO_LIBM
… …
708 #else 1354 #else
709 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e)) 1355 #define ecb_ldexpf(x,e) (float) ldexp ((double) (x), (e))
710 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e)) 1356 #define ecb_frexpf(x,e) (float) frexp ((double) (x), (e))
711 #endif 1357 #endif
712 1358
713 /* converts an ieee half/binary16 to a float */
714 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
715 ecb_function_ ecb_const float
716 ecb_binary16_to_float (uint16_t x)
717 {
718 int e = (x >> 10) & 0x1f;
719 int m = x & 0x3ff;
720 float r;
721
722 if (!e ) r = ecb_ldexpf (m , -24);
723 else if (e != 31) r = ecb_ldexpf (m + 0x400, e - 25);
724 else if (m ) r = ECB_NAN;
725 else r = ECB_INFINITY;
726
727 return x & 0x8000 ? -r : r;
728 }
729
730 /* convert a float to ieee single/binary32 */ 1359 /* convert a float to ieee single/binary32 */
731 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x); 1360 ecb_function_ ecb_const uint32_t ecb_float_to_binary32 (float x);
732 ecb_function_ ecb_const uint32_t 1361 ecb_function_ ecb_const uint32_t
733 ecb_float_to_binary32 (float x) 1362 ecb_float_to_binary32 (float x)
734 { 1363 {
… …
865 #endif 1494 #endif
866 1495
867 return r; 1496 return r;
868 } 1497 }
869 1498
870#endif 1499 /* convert a float to ieee half/binary16 */
1500 ecb_function_ ecb_const uint16_t ecb_float_to_binary16 (float x);
1501 ecb_function_ ecb_const uint16_t
1502 ecb_float_to_binary16 (float x)
1503 {
1504 return ecb_binary32_to_binary16 (ecb_float_to_binary32 (x));
1505 }
871 1506
872#endif 1507 /* convert an ieee half/binary16 to float */
1508 ecb_function_ ecb_const float ecb_binary16_to_float (uint16_t x);
1509 ecb_function_ ecb_const float
1510 ecb_binary16_to_float (uint16_t x)
1511 {
1512 return ecb_binary32_to_float (ecb_binary16_to_binary32 (x));
1513 }
873 1514
1515#endif
1516
1517#endif
1518

Diff Legend

– Removed lines
+ Added lines
< Changed lines
> Changed lines